34571 - Concatenation of string variables slow compared to strings themselves

Status: VERIFIED INVALID

(Core :: JavaScript Engine, defect, P3)

Description

[Morten Wang]

I'll be uploading a test case with various tests of string concatenation.  
Mozilla (build 2000040308) shows good performance with all the ones that uses 
strings directly, e.g. "string1" + "string2".  it's the last three it has 
problems with, they use string variables (e.g. var1 + var 2) in the 
concatenation.

try it out for yourselves.  all numberical values shown in the form fields is 
the execution time in millseconds.  the four tests on the left hand side, and 
the 2 at the top on the right hand side finished in around 1650ms on my P3/450. 
 this is just the same speed as Netscape Comm 4.72.  On the last three tests on 
the right hand side NC4.72 uses 7000ms, 10000ms and around 4500ms respectively, 
while Mozilla suddenly uses 10000ms, 14750ms and 5500ms on the same three tests. 
 I'm slightly surprised by this sudden large increase in execution time.

the test results are very positive compared to IE5.01 though, except for the 
three tests with variables in them.  the 4 tests on the left hand side, from top 
to bottom, finish in around 5.5s, 9s, 12.5s and 16s in IE5.01.  in other words, 
a nearly linear increase in usage for each string that gets added.  the two top 
tests on the right hand side finish in around 9.3s and 20s, a _huge_ difference 
from both Mozilla and Communicator.  the last three tests, with variables, 
execute at just about the same speed as Communicator though (the last one 
actually about a second faster).

Comment 1

[Morten Wang]

Attachment: String concatenation test page

Comment 2

[Morten Wang]

adding keyword 'testcase' as per bugathon guidelines

Comment 3

[rogerl (gone)]

--> phil, would you try running against the shell to see if this is embedding 
related or an engine specific issue?

Comment 4

[Gervase Markham [:gerv]]

Confirming. The last three cases currently crash Mozilla - bug 31925 filed.

Gerv

Comment 5

[Phil Schwartau]

Note - there is a transposition above: the related bug is 39125, not 31925.

Comment 6

[Phil Schwartau]

Thank you for these tests !!!

I have checked this issue with the JavaScript Engine team. Concatenation of 
literal strings is faster than that of variables because the JavaScript compiler 
has been optimized. Concatenations get turned into byte code by the compiler 
something like this: 

SCRIPT				         BYTE CODE

str = 'Hello' + var         ===>         Add  'Hello', var, str	
                         (i.e. add 'Hello' and var, and put the result into str)


str = 'Hello'  + ' there'   ===>         Mov  'Hello there', str
	                           (i.e. put 'Hello there' directly into str)


The compiler is optimized to realize that in the literal concatenation, str 
cannot be anything other than 'Hello there', so it proceeds to put this value 
directly into str. If the compiler were not optimized, all additions would be 
equally slow as addition of variables … 

So this bug as stated must be marked as invalid. However, the reporter's test 
cases have been extremely helpful, and have uncovered an out-of-memory issue 
filed as bug 39125.  Also related are bug 2235 and bug 3649.

For the record, here are times in milliseconds I recorded for the reporter's own 
tests. I was testing the JavaScript from the JavaScript shell, pulled and built 
from the Mozilla cvs repository on 06/01/00, rather than from the Mozilla 
browser embedding. Note that the time growth is approximately linear as we 
increase the number of iterations:


Concatenation 	        10^3 iterations	10^4 	10^5 	10^6 
				
two small strings	        0	16	250	2453
three small strings	        0	16	250	2453
four small strings	        0	16	250	2453
five small strings	        0	16	250	2453
two large strings	        0	16	250	2453
three large strings	        0	16	250	2469
two strings and a variable	0	63	688	out of memory
two strings and two variables	0	94	969	out of memory
two variables	                0	47	469	out of memory

Comment 7

[Phil Schwartau]

Note: my OS for these results was WinNT 4.0 (SP5)

Comment 8

[Morten Wang]

Thanks for the explanation Phil.  I'm glad I could be of some help, even though 
this actual bug is invalid.  Seems like I've discovered an area where all 
browsers take a performance hit, something I'll keep in mind for later work.

Comment 9

[Keyser Sose]

Problems with browsers in general not Mozilla marking as Verified.